| 基于通道注意力机制卷积神经网络的TOF 激光雷达目标测距研究 |
| 点此下载全文 |
| 引用本文：曹明玉,桂林.基于通道注意力机制卷积神经网络的TOF 激光雷达目标测距研究[J].上海第二工业大学(中文版),2025,42(4):418-424 |
| 摘要点击次数: 169 |
| 全文下载次数: 19 |
|
|
| 中文摘要:同步定位与建图(synchronous positioning and mapping, SLAM) 技术是自动驾驶和AR/VR 等领域的关键技术之一, 而飞行时间(Time-of-Flight, TOF) 激光雷达是SLAM 技术在复杂环境下精确定位的重要组成。然而, 在复杂环境中, 各种干扰导致TOF 激光雷达的测距精度降低, 使其无法准确探测目标深度信息, 且传统方法也难以对多种干扰同时进行有效校正。针对这些问题, 提出了一种基于通道注意力机制- 二维卷积神经网络(channel attention mechanism-two-dimensional convolutional neural network, C-channel-2D-CNN) 的误差校正模型, 通过数据驱动的方法对TOF 测量中产生的系统误差和非系统误差进行统一校正。本文实验搭建基于DCAM710 商业TOF 激光雷达测距系统, 通过改变TOF 激光雷达与被测物体之间的距离, 采集了1.04.0 m 范围内的深度数据。首先, 设计并优化2D-CNN 结构, 确定了网络层数、卷积核大小等参数; 随后在2D-CNN 网络中加入了通道注意力机制, 以增强模型对环境中误差特征的关注度, 从而更好地学习各像素的深度信息; 最后通过均方根误差(root means square error, RMSE)、平均绝对误差(mean absolute error, MAE) 等作为模型评价指标, 与传统CNN 模型结果进行比较, 分析其在不同距离下的性能。实验结果表明, 本文所提的CNN 误差校正模型显著降低了TOF 激光雷达的测距误差, 校正后测距的RMSE 约为传统CNN模型的1/2。该研究为提升SLAM 系统的定位准确性提供了一种有效方案。 |
| 中文关键词:飞行时间激光雷达 卷积神经网络 通道注意力机制 误差校正 |
| |
| Research on Target Ranging of TOF Lidar Based on Channel Attention Mechanism Convolution Neural Network |
|
|
| Abstract:Synchronous positioning and mapping (SLAM) technology is one of the key technologies in the fields of autopilot and AR/VR, and Time of Flight (TOF) laser radar is an important component of SLAM technology for accurate positioning in complex environments. However, in the complex environment, various interferences lead to decreases of the ranging accuracy of the TOF laser radar, and it is unable to accurately detect the depth information of the target, and the traditional methods can not correct multiple interferences at the same time. To solve these problems, an error correction model based on the channel attention mechanism-two-dimensional convolutional neural network (C-channel-2D-CNN) is proposed, and the systematic and non-systematic errors generated in the TOF measurement are uniformly corrected through the data driven method. In this paper, a commercial TOF laser radar ranging system based on DCAM710 is built experimentally. By changing the distance between the TOF laser radar and the measured object, the depth data from the range of 1.0 to 4.0 m are collected. Firstly, the 2D-CNN structure is designed and optimized to determine the number of network layers, the size of convolution kernel and other parameters. Then, the channel attention mechanism is added to the 2D-CNN network to enhance the attention to the errors feature in the environment, and better learn the depth information of each pixel. Finally, the root means square error (RMSE), mean absolute error (MAE) and other model evaluation indicators are used to compare with the results of the traditional CNN model, and analyze its performance at different distances. The experimental results show that the CNN error correction model in this paper greatly reduces the ranging error of TOF laser radar. The RMSE of the range of this corrected model is about 1/2 of the traditional CNN model, which provides an effective method to improve the positioning accuracy of SLAM system. |
| keywords:Time-of-Flight laser radar convolutional neural network channel attention mechanism error correction |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|